Linear-Speed Vertex Cache Optimisation

Tom Forsyth, RAD Game Tools: tom.forsyth@eelpi.gotdns.org

28th September 2006

 

Introduction

This paper introduces an algorithm for optimising an indexed triangle list to be friendly to a wide variety of hardware vertex caches of unknown size. The method has the advantage of being universally good for a wide range of cache sizes and replacement policies, of running in linear (and fast) time and space proportional to the number of triangles in the mesh. It is also within a few percent performance of the best known alternative algorithms, and easy to understand and tweak if needed.

 

Background

Indexed primitives have replaced strips and fans as the most common rendering primitive in graphics hardware today. When rendering each triangle, the vertices it uses must be processed (for example, transformed to screen space and lit in various ways) before rendering. Indexed rendering hardware typically uses a small cache of vertices to avoid re-processing vertices that are shared between recently-rendered triangles. Using a moderately-sized cache (anywhere between 12 and 24 vertices is common) reduces the load on the vertex processing pipeline far more than the older non-indexed strips and fans model – often halving the amount of work required per triangle.

 

However, the vertex cache relies on the ordering of the triangles to be optimised for its characteristics. If the triangles are sent in a random order, the vertices almost always miss the cache and there is no benefit obtained. If triangles are sent in clusters so that each triangle is close to previously-rendered triangles, it is more likely that more of its vertices are still in the cache.

 

Thus, finding the correct order to render triangles in is important to reduce the work of the vertex processing pipeline. The fundamental problem is that meshes are typically two-dimensional surfaces, and caches are at their most efficient when streaming one-dimensional data. Thus the ideal ordering is far from obvious.

 

[Hoppe] discussed an algorithm that optimised the order of triangles for a given cache of a certain size and replacement policy. This is useful for a hardware target that is fixed and known, such as a games console. However, if the hardware cache is a different size or type, the algorithm needs to be re-run for that new configuration.

 

On many platforms, such as the PC or Macintosh, and also for shipping the same art for many different consoles, the exact size and type of the vertex cache is not known at authoring time, and frequently not even known at runtime, as the exact structures are often proprietary information. For this reason, [Bogomjakov] discusses a “universal rendering sequence”, a single sequence that is favourable to a wide range of cache sizes and types, and has no particular pathological weaknesses on any particular one of them.

 

The usual way to refer to the efficiency of a vertex cache is to divide the number of vertex-cache misses by the number of triangles in the mesh. This gives the average cache miss ratio (ACMR) – the average number of vertices that need to be processed per triangle.

 

Outline

The algorithm presented here is based on a cache of vertices using a least recently used (LRU) replacement policy. Various “scores” are assigned to each vertex depending on its position within the LRU cache (i.e. how recently-used it is). Triangles are also assigned a score, which is the sum of the scores of the vertices it uses. When considering which triangle in the mesh to add to the sequence next, the triangle with the greatest score is added. The three vertices for the triangle are then moved or added to the top of the cache, and the other vertices are shuffled down. The algorithm is greedy and never backtracks or looks ahead at its choices of triangle.

 

Although most hardware uses FIFO-replacement caches for simplicity and speed, here we use an LRU-replacement cache. This cache is referred to as the “modelled cache”, as opposed to the real “hardware cache” that an actual graphics card has. The reason the modelled cache is LRU rather than FIFO is that the single sequence of triangles this algorithm generates is intended to be usable on a wide range of (usually unknown) hardware cache sizes and replacement policies, and modelling with a FIFO cache of the incorrect size causes either pathological “thrashing” of the real hardware cache, or under-utilisation. In practice, an LRU modelled cache produces a more consistent sequence, even if it does not match common hardware.

 

The modelled cache size chosen merely needs to be “big enough” to include all likely hardware caches. [Hoppe] shows that it is unlikely that any hardware will see any significant benefit from a cache larger than 32 entries, and so 32 was chosen as the modelled cache size here. Experiments were also performed with a 64-entry modelled cache, but no significant difference in eventual cache performance was found – the algorithm simply ran slower.

 

As described, the algorithm is a greedy algorithm with no lookahead or backtracking. Once it has chosen to add a triangle to the rendering sequence, it keeps that decision. Although the choice of triangle affects future choices through the vertex and triangle scores, the current triangle is not chosen in order to influence those future choices. In this way, the algorithm remains very fast, and not only runs in linear time proportional to the number of triangles in the mesh, but the linear time is very fast. This speed is extremely helpful in authoring pipelines, where artists need to have fast, accurate previews of their models.

 

It is also helpful for games that have dynamically generated content, such as joining together parts of bodies and faces to make composite characters – the resulting mesh should be optimised as a whole for the hardware vertex cache for maximum efficiency, but this cannot be performed until all the parts have been joined into a single mesh. In games such as RPGs or MMOs with a massive number of potential body-shapes, each specified dynamically, new meshes can appear every second, and algorithms that take significant processing time to optimise a mesh are counterproductive.

 

Vertex Scores in Detail

The score of a given vertex indicates how likely a triangle using it will be added to the rendered list next. High scores make it more likely, low scores make it less. The score depends on a number of things.

 

The first is that the more recently the vertex was used, the higher its score. The score is calculated by taking the vertex’s position in the LRU cache as a fractional number between zero and one, with zero at the last (LRU) position and one in the first (MRU) position. This scaled position is then raised to a power greater than one. The value of the power was chosen by repeated simulation, and the value 1.5 seems to give good results. Using a power rather than a simple linear score seems to give a behaviour that is more scale-independent – a desired trait in an algorithm attempting to optimise for an unknown size of hardware cache.

 

The exception to this rule is that the most recent three vertices, i.e. those used by the most recently added triangle, have a fixed score. The reason they are position-independent is that the order in which those three vertices were used depends entirely upon the order they were specified in the most recent triangle, and that is subject to the whims of export pipelines and various other unpredictable factors. Note that the scores for the rest of the vertices in the cache is performed as if these first three were not there, i.e. the fourth vertex in the cache gets a “top” score of 1.0. The surprising twist is that the score the first three are fixed at was also found by experimentation to be best set to 0.75. This is lower than the score of the vertex after them in the cache!

 

The reasoning for this is that if the three most-recent vertices get the highest score, the next triangle to be added is almost certain to be one that shares an edge (two of the vertices) of the most recently-added triangle. On the face of it, this sounds like exactly what we want. The problem is that because this algorithm has no look-ahead capability, it will first form a long thin strip through the mesh, and then the strip will turn back on itself for a long way, and then again. Although this sounds fairly good, it is only getting an ACMR of around 1.0 – one vertex per triangle – far worse than a good algorithm is capable of. By slightly discouraging immediate reuse of an edge, the algorithm is forced to look a little further away, and this seems to produce a more generally acceptable pattern that is not unlike the theorised ideal shape – a Hilbert curve. Here is a table showing how position in the first half of the 32-entry LRU cache corresponds to score:

 

LRU pos

Score

0

0.75

1

0.75

2

0.75

3

1.00

4

0.99

5

0.98

6

0.97

7

0.95

8

0.93

9

0.91

10

0.88

11

0.86

12

0.83

13

0.80

14

0.77

15

0.73

EMERGENCY EDIT: Actually, the numbers in the table above are wrong! It's the right sort of shape, but I typed the Excel formula in badly, so the actual numbers aren't correct. However, the code listed below is correct, so keep using that.

 

Added to this vertex score is a boost to a vertex if its remaining valence is low. The remaining valence of a vertex is the number of triangles that use it that have not yet been drawn. A vertex with only one triangle left to be drawn gets the highest boost, and a vertex with a lot of triangles yet to be drawn gets a much lower boost. The increase in score is proportional to the number of triangles still remaining using it, raised to a negative fractional power (testing showed that -0.5 is most effective), and then scaled relative to the FIFO score. Again, testing showed that a scale of 2.0 was most effective. Here is a table of vertices with reasonable valences and their score:

 

Valence

Score

1

2.00

2

1.41

3

1.15

4

1.00

5

0.89

6

0.82

7

0.76

8

0.71

9

0.67

10

0.63

 

The reasoning behind rewarding low valence is that without it, the algorithm will still pick out the best path through a mesh, maintaining a low vertex/triangle ratio. The problem is that it will leave some lone or small clusters of triangles behind as it does so, because adding them is slightly less optimal than adding other triangles. By the time the mesh is 90% done, all that remains are these little bits of “detritus” – isolated fragments that still need to be drawn. Although at this point the ACMR is still pleasingly low, adding these remaining fragments is very costly, because they are not close to each other, and so each incur costs of around 3 vertices per triangle. This can have a terrible effect on the ACMR for the mesh as a whole, and drives efficiency down again. Far better is to add these bits of detritus to the rendering order as the algorithm goes along. Adding them at the time is very slightly sub-optimal, but adding them later would be very costly. Boosting the score of a vertex with very few triangles remaining encourages the addition of the detritus at a sensible time.

 

Additionally, if the algorithm runs into a “dead end” part of the mesh, where there are no more nearby triangles to add (for example, at the end of a finger of a hand), it encourages the algorithm to start again at triangle that is near another “dead end”, such as the edge of a mesh, rather than start right in the middle of a patch and create an artificial dead end.

 

The two scores – that due to the FIFO position and the one due to remaining valence – are added together to find the total vertex score. To find each triangle’s score, the score for each of the three vertices is added together. And at each stage, the triangle with the lowest score is added to the draw list, and the scores for the vertices are recalculated.

 

The full code for calculating the score of a vertex is as follows:

 

const float FindVertexScore_CacheDecayPower = 1.5f;

const float FindVertexScore_LastTriScore = 0.75f;

const float FindVertexScore_ValenceBoostScale = 2.0f;

const float FindVertexScore_ValenceBoostPower = 0.5f;

float FindVertexScore ( vcache_vertex_data *VertexData )

{

    if ( VertexData->NumActiveTris == 0 )

    {

        // No tri needs this vertex!

        return -1.0f;

    }

 

    float Score = 0.0f;

    int CachePosition = VertexData->CacheTag;

    if ( CachePosition < 0 )

    {

        // Vertex is not in FIFO cache - no score.

    }

    else

    {

        if ( CachePosition < 3 )

        {

            // This vertex was used in the last triangle,

            // so it has a fixed score, whichever of the three

            // it's in. Otherwise, you can get very different

            // answers depending on whether you add

            // the triangle 1,2,3 or 3,1,2 - which is silly.

            Score = FindVertexScore_LastTriScore;

        }

        else

        {

            assert ( CachePosition < MaxSizeVertexCache );

            // Points for being high in the cache.

            const float Scaler = 1.0f / ( MaxSizeVertexCache - 3 );

            Score = 1.0f - ( CachePosition - 3 ) * Scaler;

            Score = powf ( Score, FindVertexScore_CacheDecayPower );

        }

    }

 

    // Bonus points for having a low number of tris still to

    // use the vert, so we get rid of lone verts quickly.

    float ValenceBoost = powf ( VertexData->NumActiveTris,

                                -FindVertexScore_ValenceBoostPower );

    Score += FindVertexScore_ValenceBoostScale * ValenceBoost;

 

    return Score;

}

 

The Constants

The four constants used in the code above were determined by a process of guesswork, judgement, and a few days of brute-force annealing of each one with fairly typical data sets and the results sent through simulations of fairly typical hardware cache structures, taking the best aggregate scores. It is entirely possible that slightly better settings and values exist that have not been discovered yet. It is also entirely possible that using power functions for the scores is not the appropriate function, and that using other functions, or possibly fully custom look-up-tables, may yield better results. However, the results are fairly close to what is considered optimal, and it is likely that any remaining improvements are small. In summary, these values appear to be “good enough” for most purposes.

 

As a side note, it is curious that the four constants are all so close to “normal” numbers. The values 1.5, 0.75, 2.0 and 0.5 seem almost boring. And yet the process of annealing did consider possible values for all of them that varied in 0.05 increments, out to a range of at least +/- 0.5 from their current settings. Nor were these the values selected using human judgement at the start of the annealing process (for example, the valence boost scale began far smaller than 2.0 – it is somewhat surprising it turned out to be so important). Nevertheless, these values consistently produced the best overall scores across the range of caches and meshes.

 

Runtime Efficiency

The draw-ordering of the triangles is as given in the algorithm above. However, there are many ways to implement this algorithm that run slowly, and one of the key features of this algorithm is speed.

 

Unlike many conventional mesh topology algorithms, no edge data is kept, nor is any needed. This avoids the expense of finding edge data (typically somewhat laborious), and also the problems inherent in pathological meshes where a single edge is shared by many (which in typical algorithms means more than two!) triangles.

 

Instead, each vertex holds the following data:

Note that the total number of triangles using the vertex is not strictly necessary for the algorithm to run, but is useful for debugging. The list of triangle indices is always this length, but it is ordered so the triangle indices yet to be added are listed first, followed by the triangle indices that have already been added to the draw list.

 

Each triangle in the mesh also stores the following data:

 

The other data structure is the FIFO cache itself, which is simply a list of vertex indices ordered from most recently used to least. Notionally this cache is 32 entries long, though for ease of implementation it actually grows to three entries longer.

 

Initialisation of the data is fairly straightforward, with two passes over the triangle data. The first simply increments the counter of the number of triangles that use each vertex, the second allocates the lists of triangles for each vertex and fills them in. After that, the score of each vertex is found using the above code, and then the score of each triangle is found by summing the scores of each vertex the triangle uses. These scores are simply cached values of the computed scores, and are updated when necessary as the algorithm runs.

 

Then comes the main body of the algorithm, which picks one triangle at a time to add to the list of drawn triangles, until there are no more triangles left to draw. Usually, the algorithm knows from the previous iteration which triangle has the highest score, and simply picks that one. In some cases, for example the first time the algorithm runs, it does not already know the best triangle, or the supposed best triangle has an unusually low score (implying that there may be others in the mesh that are better). In those rare circumstances, it runs through all the remaining triangles in the mesh searching for the best score.

 

[HINDSIGHT: Actually, the "unusually low score" clause turns out to have almost no impact on the results. It was useful while I was developing the algorithm, but it turns out that for the final algorithm presented here, it happens for only three triangles in my entire set of test data, so it really has no significant effect on the efficiency.]

 

The best triangle is then added to the draw list. For each vertex the triangle used, the valence of that vertex (the number of triangles not yet drawn that use it) is reduced by one, and the list of triangle indices in the vertex is updated appropriately.

 

The three vertices used by the triangle are either moved to the head of the LRU modelled cache, or added to the head if they were not already in it. The cache is temporarily grown in size by three vertices to include all vertices that were previously in the cache, and up to the three new ones of this triangle. Then the new positions of the vertices in the cache are updated, their corresponding scores are found using the code given above, and the scores of all their still-to-be-added triangles are also updated.

 

As this is done, the score and index of the highest-scoring triangle so far are noted. Although only the triangles that use vertices currently in the cache are scanned for their score, this is a reasonable estimate of the best score of any triangle in the entire mesh. Debugging code that exhaustively checks whether this estimate is true found it to be false only a few times in all the testing done, and even then the score difference between the correct answer and the approximation was very small. In practice, this difference is unlikely to make any significant difference to the cache-performance, and hugely decreases the runtime cost of the algorithm.

 

Finally, the cache is shrunk back to its normal size, with up to three vertices falling out of it. Their cache positions have already been updated appropriately. The algorithm repeats until there are no triangles left to add.

 

Examining the above carefully, we can see that with a “normal” mesh where each vertex is used by around 6 triangles, the runtime speed and memory use of this algorithm is linearly proportional to the number of vertices in the mesh. This is considerably better than various look-ahead or hierarchical subdivision methods, or those that use complex topological information.

 

Results

A variety of test meshes were used during the development of this algorithm. Results for four of them are shown here. Three of them are representative of typical interactive (i.e. games) meshes, and one is a more idealised mesh used for sanity-checking the algorithm. The results for the idealised mesh are still relevant for highly regular surfaces such as terrain height-fields or finely-tessellated models, but the results from it did not guide the direction of the algorithm significantly – far more importance was given to the more realistic meshes.

 

The meshes are as follows:

 

The algorithm was performed on each mesh using a modelled cache of 32 entries. 64 entries was also tried, but it appeared to make little difference to the output, it simply made the algorithm run slower. The single output was then fed to an idealised FIFO cache of various sizes, and the ACMR is shown in the table below. The 1024-entry FIFO cache is intended as an upper limit of the “perfect cache”, and indeed for a mesh with chunks fewer than 1024 vertices, it does indeed represent that perfect value, whatever the rendering order. Real hardware typically has between 12 and 24 entries, with 32 regarded by many as the upper limit of a useful cache before the extra entries give diminishing returns on extra hardware area.

 

The 4-entry cache is intended to mimic an older piece of hardware that was designed with traditional triangle and quad strips in mind. An ACMR of 1.0 represents the best score traditional non-indexed strips can produce, and we can see that the results here are considerably worse than that. However, it was found that any attempt to increase the efficiency for this sort of hardware significantly hurt the efficiency of all the other cache sizes. As hardware like this is now considered so old and slow in other respects as to be irrelevant, the numbers are left here for illustration only. If this hardware is still a specific target for some reason (for example, some older and portable consoles still use non-indexed rendering), it is better to run one of the algorithms custom-designed for efficient stripping of non-indexed meshes.

 

 

Hardware cache size

4

8

12

16

20

24

28

32

1024

GrannyRocks

1.556

0.908

0.820

0.797

0.781

0.776

0.771

0.765

0.732

RaybackSplit

1.559

0.893

0.791

0.769

0.755

0.747

0.738

0.730

0.695

RaybackWhole

1.596

0.874

0.756

0.727

0.713

0.703

0.691

0.682

0.647

Sphere

1.785

0.888

0.740

0.715

0.702

0.691

0.681

0.672

0.529

 

Note that the ACMR consistently decreases as more entries are added to the cache. It is important to remember that the same rendering sequence was used for all the cache efficiency measurements – the algorithm was not re-run for each cache, and indeed has no idea what size of hardware cache it will be running on. As can be seen, the ACMRs consistently and smoothly decrease, with no preference shown to any particular cache size, and no pathological cases. This is in contrast to other algorithms that need to know their target cache size. If the guess is too large, the algorithm can thrash a smaller cache, leading to terrible performance, and if the guess is too small, the algorithm shows almost no advantage from the larger cache size.

 

It is interesting to note that as predicted by Hoppe, a 32-entry cache gets very close to the performance of the (impractically large) 1024-entry cache. In most cases the improvement from 16 entries to 32 is as large as the improvement from 32 to 1024. It is reassuring to see that this algorithm produces an ordering that obtains something very close to this near-perfect ACMR.

 

The exception is the extremely regular sphere tessellation, where there is a significant difference between 32 entries and 1024. This may be because the other test meshes simply do not have enough vertices to show the full benefit of the extra cache size (RaybackWhole is 5500 vertices in size, but has quite a few seams in it from texturing constraints), or because the extremely regular nature of the mesh allows the larger cache to work even better. It is believed from experience that most real-world game meshes do not share these characteristics. Although the perfect ACMR for an infinitely large plane of triangles and a perfect cache is 0.5 vertices per triangle, real meshes do not have this regularity, and contain many seams and edges and other discontinuities that limit how well any cache can possibly perform. Nevertheless, it would be useful to obtained results from some larger real-world meshes to check these assumptions.

 

No apples-to-apples comparison with other vertex-cache methods has been made yet. Hopefully this will be done in future. However, eyeball comparison of the results in [Bogomjakov] for cache sizes of 32 entries and using the latter two meshes shows both algorithms to produce ACMRs somewhere around 0.68. The first two meshes shown here have mesh chunks much smaller than shown in the other papers referenced. Smaller chunks have more edges, and thus a worse best-possible-case (indeed, for GrannyRocks, the 1024-entry cache can fit the whole of each mesh chunk, so no rendering sequence can possibly perform better than an ACMR of 0.732)

 

Future Work

Not much. The difference between the results in this algorithm, the perfect results of the 1024-entry cache, and the best results of [Hoppe] and [Bogomjakov] are very small in practice. Eyeball estimates are that the three algorithms running on the same hardware cache span an ACMR difference of 10% at most between best and worst. While optimising vertex cache use to a state better than a random ordering or that of simple triangle strips is important, at a certain point it is eclipsed by other factors in the rendering pipeline. A 10% difference in ACMR is typically going to be dwarfed by other independent factors such as triangle setup and fillrate, and uninteresting for most practical purposes.

 

So this problem is pretty close to being “solved”. In fact I think it was solved for known hardware by [Hoppe], and for unknown hardware by [Bogomjakov], with the one weakness of the latter being the complexity of implementation and the runtime and memory costs of the algorithm. This much simpler algorithm solves those problems. The extra speed improves artist iteration time, and also allows dynamically-generated data to be optimised on the fly.

 

Another interesting version is the K-Cache algorithm [Lin]. This is a more generalised caching algorithm, but when applied to the problem of ACMR and tweaked a bit by Jason Hughes, the implementation seems to boil down to pretty much the same algorithm as outlined here. Which is reassuring in a “convergent evolution” sort of way.

 

Additional Notes

To fully optimise for rendering, there are two or three steps to take in total. The first is the optimisation for the post-transform vertex cache, as described above.

 

Next, the order of the vertices in the vertex buffer is found. Start with an empty vertex buffer. Scan through the triangles in the order found above. Check whether each of the three vertices has been added to the VB yet. If not, add them, and remember the remapping from the old index number to the new one. Keep going until you run out of triangles. Finally, remap all the old indices to the new ones, but do not change the order the triangles are rendered in. The vertices have now been reordered so that when the graphics card renders the triangles, it will read vertices in a mostly-linear fashion (it would be completely linear, except for vertices that dropped out of the post-transform vertex cache). This reordering of vertex data helps the pre-transform cache to use the memory bandwidth effectively. This step can produce just as much of a speed improvement as the reordering of the triangles, so although the algorithm is simple, it is important.

 

An optional third step is to convert the indexed triangle list that we have been using so far into some other format. The two sensible alternatives are an indexed triangle strip, and an indexed triangle strip with restarts. The difference between the two is that the latter has a special index number (usually -1) that says the strip should be restarted, without the hack of inserting degenerate triangles. There are plenty of tutorials available on how to do both forms of strips but it is important to note that you should not change the order of the triangles! You've already optimised for the vertex cache, and that is still the most important thing. Choosing between these formats is simply a matter of which one is best for your target hardware.

 

In my experience, the theoretical win from indexed triangle strips over lists is exactly that - theoretical. In practice on the Xbox1 (where the advantage is clearly understood and documented) I have never seen a speed improvement of strips with degenerates versus lists, though others claim a very slight improvement. It seems not to be worth the extra hassle of dealing with indexed strips compared to indexed lists. Depending on the mesh, one can give a slightly smaller index buffer than the other, but the difference is fairly trivial, and it is very dependent on the mesh chosen. Although strips use only one index per triangle instead of three, you need to insert so many degenerate triangles that it all seems to cancel out the advantages.

 

However, indexed triangle strips with a restart index do seem to produce significantly smaller index buffers. Instead of having to add degenerate triangles, you simply add a -1 index and the strip restarts in the new place. Tests on the platforms that support them (some consoles and DX10 cards) don't seem to show a significant speed increase so far, but the smaller index-buffer size is always useful.

 

References

[Hoppe] “Optimization of mesh locality for transparent vertex caching.” Hugues Hoppe, ACM SIGGRAPH 1999, 269-276, http://research.microsoft.com/~hoppe/

 

[Bogomjakov] “Universal rendering sequences for transparent vertex caching of progressive meshes” Alexander Bogomjakov, Craig Gotsman, Computer Graphics Forum 21(2):137-148, 2002, http://www.cs.technion.ac.il/~gotsman/publications.html

 

[Lin] “An Improved Vertex Caching Scheme for 3D Mesh Rendering” Gang Lin, Thomas P.-Y. Yu, IEEE Transactions on Visualization and Computer Graphics Volume 12, Issue 4 (July 2006) http://www.ecse.rpi.edu/~lin/K-Cache-Reorder/

Implementations and Source Code

Apologies for not posting my own implementation, but it's part of Granny3D, so while I can talk about the algorithm, posting the actual source is not cool. However, others have cleanroomed versions and made them available. However please note I have not actually checked these for correctness or bugs!

 

Martin Storsjö has source of a very clean implementation of this algorithm as part of his Master's thesis Efficient Triangle Reordering for Improved Vertex Cache Utilisation in Realtime Rendering, which is well worth a read. Thanks Martin!

 

Pat Wilson also has a good-looking version here: triListOpt.h and triListOpt.cpp. Thanks Pat!

 

Another good implementation in the source for Farbrausch's Werkzeug3. Thanks ryg!

 

Elegant standalone version by Adrian Stone. Bonus points for the "namespace Forsyth": https://github.com/bkaradzic/bgfx/tree/master/3rdparty/forsyth-too. Thanks to Branimir Karadžic.

 

 

Back to all papers.